Development and validation of a fully automated transformer-based 3D framework for pancreatic fat quantification in pancreatic steatosis

Abstract

Purpose: To develop and validate a fully automated TransUNet-based framework for 3D pancreatic segmentation and volumetric fat quantification on PDFF MRI. Specifically, to evaluate the model's performance against state-of-the-art CNN architectures and assess its clinical reproducibility compared to manual reference methods. Methods: This retrospective study involved 140 adults with metabolic dysfunction-associated steatotic liver disease who underwent 3.0 T multi-echo mDIXON MRI. A TransUNet-based model was developed using a 5-fold cross-validation approach, integrating convolutional and transformer layers to capture global and local features. Various architectures (UNet, nnUNet) and multiple input combinations of water, fat, and R2* maps were systematically compared. Segmentation performance was primarily assessed using the Dice Similarity Coefficient (DSC), with the Jaccard index, precision, recall, and 95th-percentile Hausdorff distance additionally used for network comparison. Agreement of pancreatic fat quantification across measurement methods was assessed using Bland-Altman analysis and intraclass correlation coefficients (ICC). Results: TransUNet achieved the highest segmentation accuracy (DSC: 85.46%, IQR: 80.76–86.24%), outperforming nnUNet (84.24%) and UNet (71.89%). Optimal performance was reached using water-only and fat-only input series. The AI-based volumetric method demonstrated strong agreement with manual whole-organ PDFF (r = 0.866, p < 0.001) with a minimal mean bias (1.20). Conversely, 2D ROI methods significantly underestimated pancreatic fat (6.08 ± 2.74%) compared to both manual (21.95 ± 6.14%) and AI-based (20.72 ± 6.13%) volumetric assessments (p < 0.001). Conclusion: TransUNet provides accurate, reproducible 3D pancreatic segmentation and fat quantification. By capturing the entire organ volume, this automated framework overcomes the sampling bias inherent in traditional 2D ROI methods, offering a fast and reliable biomarker for pancreatic steatosis.